deslop

Outcome: the text says the same things in plainer words with no AI tells. Every fact, command, path, link target, placeholder, and required phrase survives, and the project's own checks still pass.

Two modes, from no-ai-slop. Audit reports tells with file:line and a proposed fix and changes nothing; use it when asked to check, scan, or audit, and as the first pass over UI copy in code. Edit applies the rewrite; use it when asked to fix, rewrite, simplify, humanize, or deslop. An audit ends with an offer to edit.

When available, use three helpers: the humanizer skill (the rewrite), the no-ai-slop skill's eval.md (the check), and Vale through scripts/lint.sh (the mechanical pass). scripts/tells.sh folds stop-slop's quick checks into one grep. Load humanizer and no-ai-slop when the steps below name them; their SKILL.md files stay the source of truth for the patterns. Resolve bundled scripts from this skill's directory and target paths from the destination repository. Check helper availability first. If a helper is missing, continue direct prose review against the user's voice and project rules, and name its check as unrun. Do not install tools or invent external evaluation results. A cold Vale cache can run vale sync and download style packages; inspect that setup and its destinations before execution, within the session's existing authority.

Procedure

1. Scope. List the target files and read each one in full. Read applicable repository instructions and record existing changes. Keep edits within the requested files or copy surface. Before touching prose, record three things:
   • Required phrases and anchors. Grep the repository's tests and docs for exact phrases the target contains (grep -rn "same task" tests docs) and for inbound links to its headings (grep -rn "README.md#" docs). A test that greps prose fails on a line wrap, so keep those phrases whole.
   • Untouchable content: code blocks, inline code, commands, paths, YAML frontmatter, data, link targets. In source files, everything outside string literals, plus the placeholders inside them; read references/ui-copy.md for that branch.
   • Register. Technical, reference, and error text stays neutral and plain. Product and marketing copy keeps the writer's voice. A writing sample from the user overrides both.
2. Mechanical pass. Run available scripts/tells.sh FILES on every target, and available scripts/lint.sh FILES on Markdown and text targets. Keep the output as the before list. In audit mode, this plus a read for the patterns tells.sh cannot grep (forced triads, repeated openings, one-line closers) is the report.
3. Rewrite with available humanizer in file mode, or direct prose judgment: mark tells strongest first, draft, check the five survivors (a not-X-but-Y contrast, a one-line closer, a dash, a triad, a bold label), then write the final text. A bold label earns its place when it states a claim ("Keys stay private"), not when it names a topic followed by fragments. Do not swap tells for telegraphic fragments, dropped articles or verbs, or arrows standing in for sentences. Prose changes only. For UI copy, apply references/ui-copy.md alongside.
4. Check with available no-ai-slop: answer each item in its eval.md pass or fail. A fail means fix the draft and check again. If absent, reread for clear language and preserved facts, and report a direct check without claiming its external rubric passed.
5. Re-run step 2. Give each remaining hit a disposition: fixed, or kept with the reason. Reasons that hold: a domain or product word, a quoted example, a claim-form heading or label ("Releases are verified"), a state adjective ("is enabled"), an identifier in code. Vale's Readability grades are advisory for technical writing.
6. Verify. Run required project gates and relevant existing checks. Preserve snapshot/test update gates and render changed screens where applicable. Re-grep the required phrases and anchors from step 1. Confirm git diff shows prose-only changes.
7. Report. What changed, grouped by tell; before and after counts from the scripts when scans completed; dispositions for kept hits; each check run with its result; any check you could not run, named as unrun.

Completion

Done when every available helper check passes or its remaining hits carry a disposition, required project gates pass, and you have delivered the report. Unavailable helpers stay unrun, with a direct-review result and its limits; missing or failed scans leave their counts unavailable, never zero. In audit mode, done when the report lists every hit with file, line, quoted text, the tell, and a proposed fix, and offers the edit.

Scripts

• scripts/tells.sh FILE... greps the strongest tells in any text or source file and prints file:line hits grouped by tell. Exit 1 on a strong hit, 0 otherwise, 2 on usage or a scan failure. Strong tells justify an edit on one sighting; weak tells need company from other tells in the same passage.
• scripts/lint.sh [--all] FILE.md... runs Vale with write-good, proselint, and Readability from a cache directory, so no repository gains a config file. E-Prime is off because the humanizer prefers is, are, and has; spelling is off because technical writing is full of product and code words. Exit 2 on setup failure, with an install hint when Vale is absent (brew install vale).